Failing to convert numpy array to JSON

Question

I am trying to write a JSONEncoder so I can dump some of my Python object data to JSON text to be read in by another application. I am having particular trouble converting the numpy arrays. I have read many of the SO articles on the subject, and I have borrowed heavily from @tlausch's solution in this article SimpleJSON and NumPy array - but I have not been able to get it working.

I have a pretty simple class that demonstrates my problem:

import numpy as np
class Movement:
 def __init__(self, t1,t2,t3,t4):
 self._t1 = t1
 self._t2 = t2
 self._t3 = t3
 self._t4 = t4
 self._a0 = np.array([0,0,0])

Here is the code from my JSONEncoder:

import json
import base64
import numpy as np
class MovementEncoder(json.JSONEncoder):
 def default(self, obj):
 if isinstance(obj, np.ndarray):
 if obj.flags['C_CONTIGUOUS']:
 obj_data = obj.data
 else:
 cont_obj = np.ascontiguousarray(obj)
 assert(cont_obj.flags['C_CONTIGUOUS'])
 obj_data = cont_obj.data
 data_b64 = base64.b64encode(obj_data)
 return dict(__ndarray__ = data_b64, dtype = str(obj.dtype), shape = obj.shape)
 try:
 my_dict = obj.__dict__ ## <-- ERROR raised here
 except TypeError:
 pass
 else:
 return my_dict
 return json.JSONEncoder.default(self, obj)

When I create a simple Movement object, and then call encoder.encode(obj), I get the following error:

>>> obj = Movement(1,2,3,4)
>>> encoder = MovementEncoder()
>>> encoder.encode(obj)
...
 'bytes' object has no attribute '__dict__'

By adding some print statements to the encoder, I can see that it is correctly recursing from the object, to the object's dictionary, then to the attribute that has the np.array type. I thought the point of this solution was that the base64 representation of the ndarray type was JSON encodable by default, but that appears not to be the case. Where did I go wrong?

Note: using Python 3.4, and NumPy 1.8.2

EDIT: updated code to show where error occurs

• ForeverWintr: Your code runs successfully for me on python 2.7.5 with numpy 1.7.1
• gariepy: Oh, forgot to specify python 3.4 and numpy 1.8.2...
• Alex Hall: Based on the full traceback, which line is throwing that error?
• gariepy: @AlexHall Updated Question to point to where error occurs
• Alex Hall: Why not catch AttributeError instead of TypeError?

Answer 1

I was finally able to solve this problem by adjusting the return value for numpy ndarray types to return a list, instead of a dict. I used the numpy builtin method tolist(), which returns a json-encodable representation of an ndarray.

import json
import base64
import numpy as np
class MovementEncoder(json.JSONEncoder):
 def default(self, obj):
 if isinstance(obj, np.ndarray):
 if obj.flags['C_CONTIGUOUS']:
 obj_data = obj.data
 else:
 cont_obj = np.ascontiguousarray(obj)
 assert(cont_obj.flags['C_CONTIGUOUS'])
 obj_data = cont_obj.data
 ## data_b64 = base64.b64encode(obj_data)
 ## converting to base64 and returning a dictionary did not work
 ## return dict(__ndarray__ = data_b64, dtype = str(obj.dtype), shape = obj.shape)
 return obj.tolist() ## instead, utilize numpy builtin tolist() method
 try:
 my_dict = obj.__dict__ ## <-- ERROR raised here
 except TypeError:
 pass
 else:
 return my_dict
 return json.JSONEncoder.default(self, obj)